CVPR 2017 Papers

This paper presents a material classification method using an off-the-shelf Time-of-Flight (ToF) camera. We use a key observation that the depth measurement by a ToF camera is distorted in objects with certain materials, especially with translucent materials. We show that this distortion is caused by the variations of time domain impulse responses across materials and also by the measurement mechanism of the existing ToF cameras. Specifically, we reveal that the amount of distortion varies according to the modulation frequency of the ToF camera, the material of the object, and the distance between the camera and object. Our method uses the depth distortion of ToF measurements as features and achieves material classification of a scene. Effectiveness of the proposed method is demonstrated by numerical evaluation and real-world experiments, showing its capability of even classifying visually similar objects.

Problems at the intersection of vision and language are of significant importance both as challenging research questions and for the rich set of applications they enable. However, inherent structure in our world and bias in our language tend to be a simpler signal for learning than visual modalities, resulting in models that ignore visual information, leading to an inflated sense of their capability. We propose to counter these language priors for the task of Visual Question Answering (VQA) and make vision (the V in VQA) matter! Specifically, we balance the popular VQA dataset (Antol et al., ICCV 2015) by collecting complementary images such that every question in our balanced dataset is associated with not just a single image, but rather a pair of similar images that result in two different answers to the question. Our dataset is by construction more balanced than the original VQA dataset and has approximately twice the number of image-question pairs. Our complete balanced dataset is publicly available at http://visualqa.org/ as part of the 2nd iteration of the Visual Question Answering Dataset and Challenge (VQA v2.0). We further benchmark a number of state-of-art VQA models on our balanced dataset. All models perform significantly worse on our balanced dataset, suggesting that these models have indeed learned to exploit language priors. This finding provides the first concrete empirical evidence for what seems to be a qualitative sense among practitioners. Finally, our data collection protocol for identifying complementary images enables us to develop a novel interpretable model, which in addition to providing an answer to the given (image, question) pair, also provides a counter-example based explanation. Specifically, it identifies an image that is similar to the original image, but it believes has a different answer to the same question. This can help in building trust for machines among their users.

We present a theoretically grounded approach to train deep neural networks, including recurrent networks, subject to class-dependent label noise. We propose two procedures for loss correction that are agnostic to both application domain and network architecture. They simply amount to at most a matrix inversion and multiplication, provided that we know the probability of each class being corrupted into another. We further show how one can estimate these probabilities, adapting a recent technique for noise estimation to the multi-class setting, and thus providing an end-to-end framework. Extensive experiments on MNIST, IMDB, CIFAR-10, CIFAR-100 and a large scale dataset of clothing images employing a diversity of architectures --- stacking dense, convolutional, pooling, dropout, batch normalization, word embedding, LSTM and residual layers --- demonstrate the noise robustness of our proposals. Incidentally, we also prove that, when ReLU is the only non-linearity, the loss curvature is immune to class-dependent label noise.

360deg video requires human viewers to actively control "where" to look while watching the video. Although it provides a more immersive experience of the visual content, it also introduces additional burden for viewers; awkward interfaces to navigate the video lead to suboptimal viewing experiences. Virtual cinematography is an appealing direction to remedy these problems, but conventional methods are limited to virtual environments or rely on hand-crafted heuristics. We propose a new algorithm for virtual cinematography that automatically controls a virtual camera within a 360deg video. Compared to the state of the art, our algorithm allows more general camera control, avoids redundant outputs, and extracts its output videos substantially more efficiently. Experimental results on over 7 hours of real "in the wild" video show that our generalized camera control is crucial for viewing 360deg video, while the proposed efficient algorithm is essential for making the generalized control computationally tractable.

Multi-instance multi-label (MIML) learning has many interesting applications in computer visions, including multi-object recognition and automatic image tagging. In these applications, additional information such as bounding-boxes, image captions and descriptions is often available during training phrase, which is referred as privileged information (PI). However, as existing works on learning using PI only consider instance-level PI (privileged instances), they fail to make use of bag-level PI (privileged bags) available in MIML learning. Therefore, in this paper, we propose a two-stream fully convolutional network, named MIML-FCN+, unified by a novel PI loss to solve the problem of MIML learning with privileged bags. Compared to the previous works on PI, the proposed MIML-FCN+ utilizes the readily available privileged bags, instead of hard-to-obtain privileged instances, making the system more general and practical in real world applications. As the proposed PI loss is convex and SGD-compatible and the framework itself is a fully convolutional network, MIML FCN+ can be easily integrated with state-of-the-art deep learning networks. Moreover, the flexibility of convolutional layers allows us to exploit structured correlations among instances to facilitate more effective training and testing. Experimental results on three benchmark datasets demonstrate the effectiveness of the proposed MIML-FCN+, outperforming state-of-the-art methods in the application of multi-object recognition.

The inability to interpret the model prediction in semantically and visually meaningful ways is a well-known shortcoming of most existing computer-aided diagnosis methods. In this paper, we propose MDNet to establish a direct multimodal mapping between medical images and diagnostic reports that can read images, generate diagnostic reports, retrieve images by symptom descriptions, and visualize attention, to provide justifications of the network diagnosis process. MDNet includes an image model and a language model. The image model is proposed to enhance multi-scale feature ensembles and utilization efficiency. The language model, integrated with our improved attention mechanism, aims to read and explore discriminative image feature descriptions from reports to learn a direct mapping from sentence words to image pixels. The overall network is trained end-to-end by using our developed optimization strategy. Based on a pathology bladder cancer images and its diagnostic reports (BCIDR) dataset, we conduct sufficient experiments to demonstrate that MDNet outperforms comparative baselines. The proposed image model obtains state-of-the-art performance on two CIFAR datasets as well.

Prior approaches to line segment detection typically involve perceptual grouping in the image domain or global accumulation in the Hough domain. Here we propose a probabilistic algorithm that merges the advantages of both approaches. In a first stage lines are detected using a global probabilistic Hough approach. In the second stage each detected line is analyzed in the image domain to localize the line segments that generated the peak in the Hough map. By limiting search to a line, the distribution of segments over the sequence of points on the line can be modeled as a Markov chain, and a probabilistically optimal labelling can be computed exactly using a standard dynamic programming algorithm, in linear time. The Markov assumption also leads to an intuitive ranking method that uses the local marginal posterior probabilities to estimate the expected number of correctly labelled points on a segment. To assess the resulting Markov Chain Marginal Line Segment Detector (MCMLSD) we develop and apply a novel quantitative evaluation methodology that controls for under- and over-segmentation. Evaluation on the YorkUrbanDB dataset shows that the proposed MCMLSD method outperforms the state-of-the-art by a substantial margin.

Learning robust regression model from high-dimensional corrupted data is an essential and difficult problem in many practical applications. The state-of-the-art methods have studied low-rank regression models that are robust against typical noises (like Gaussian noise and out-sample sparse noise) or outliers, such that a regression model can be learned from clean data lying on underlying subspaces. However, few of the existing low-rank regression methods can handle the outliers/noise lying on the sparsely corrupted disjoint subspaces. To address this issue, we propose a low-rank-sparse subspace representation for robust regression, hereafter referred to as LRS-RR in this paper. The main contribution include the following: (1) Unlike most of the existing regression methods, we propose an approach with two phases of low-rank-sparse subspace recovery and regression optimization being carried out simultaneously;(2) we also apply the linearized alternating direction method with adaptive penalty to solved the formulated LRS-RR problem and prove the convergence of the algorithm and analyze its complexity; (3) we demonstrate the efficiency of our method for the high-dimensional corrupted data on both synthetic data and two benchmark datasets against several state-of-the-art robust methods.

Zero-shot learning for visual recognition has received much interest in the most recent years. However, the semantic gap across visual features and their underlying semantics is still the biggest obstacle in zero-shot learning. To fight off this hurdle, we propose an effective Low-rank Embedded Semantic Dictionary learning (LESD) through ensemble strategy. Specifically, we formulate a novel framework to jointly seek a low-rank embedding and semantic dictionary to link visual features with their semantic representations, which manages to capture shared features across different observed classes. Moreover, ensemble strategy is adopted to learn multiple semantic dictionaries to constitute the latent basis for the unseen classes. Consequently, our model could extract a variety of visual characteristics within objects, which can be well generalized to unknown categories. Extensive experiments on several zero-shot benchmarks verify that the proposed model can outperform the state-of-the-art approaches.

Pooling second-order local feature statistics to form a high-dimensional bilinear feature has been shown to achieve state-of-the-art performance on a variety of fine-grained classification tasks. To address the computational demands of high feature dimensionality, we propose to represent the covariance features as a matrix and apply a low-rank bilinear classifier. The resulting classifier can be evaluated without explicitly computing the bilinear feature map which allows for a large reduction in the compute time as well as decreasing the effective number of parameters to be learned. To further compress the model, we propose a classifier co-decomposition that factorizes the collection of bilinear classifiers into a common factor and compact per-class terms. The co-decomposition idea can be deployed through two convolutional layers and trained in an end-to-end architecture. We suggest a simple yet effective initialization that avoids explicitly first training and factorizing the larger bilinear classifiers. Through extensive experiments, we show that our model achieves state-of-the-art performance on several public datasets for fine-grained classification trained with only category labels. Importantly, our final model is an order of magnitude smaller than the recently proposed compact bilinear model [??], and three orders smaller than the standard bilinear CNN model [??].

In this work we introduce a novel loss max-pooling concept for handling imbalanced training data distributions, applicable as alternative loss layer in the context of deep neural networks for semantic image segmentation tasks. Most real-world semantic segmentation datasets exhibit long tail distributions with few object categories comprising the majority of data and consequently biasing the classifiers towards them. Our method adaptively re-weights the contributions of each pixel based on their observed losses, targeting under-performing classification results as often encountered for under-represented object classes. Moreover, our approach goes beyond conventional cost-sensitive learning attempts through adaptive considerations that allow us to indirectly address both, inter- and intra-class imbalances. We provide a theoretical justification of our approach, complementary to experimental analyses on standard semantic segmentation datasets. In our experiments on the challenging Cityscapes and Pascal VOC 2012 segmentation benchmarks we find consistently improved results, demonstrating the efficacy of our approach.

Human parsing has recently attracted a lot of research interests due to its huge application potentials. However existing datasets have limited number of images and annotations, and lack the variety of human appearances and the coverage of challenging cases in unconstrained environment. In this paper, we introduce a new benchmark "Look into Person (LIP)" that makes a significant advance in terms of scalability, diversity and difficulty, a contribution that we feel is crucial for future developments in human-centric analysis. This comprehensive dataset contains over 50,000 elaborately annotated images with 19 semantic part labels, which are captured from a wider range of viewpoints, occlusions and background complexity. Given these rich annotations we perform detailed analysis of the leading human parsing approaches, gaining insights into the success and failures of these methods. Furthermore, in contrast to the existing efforts on improving the feature discriminative capability, we solve human parsing by exploring a novel self-supervised structure-sensitive learning approach, which imposes human pose structures into parsing results without resorting to extra supervision (i.e., no need for specifically labeling human joints in model training). Our self-supervised learning framework can be injected into any advanced neural networks to help incorporate rich high-level knowledge regarding human joints from a global perspective and improve the parsing results. Extensive evaluations on our LIP and the public PASCAL-Person-Part dataset demonstrate the superiority of our method.

Recognizing fine-grained categories (e.g., bird species) is difficult due to the challenges of discriminative region localization and fine-grained feature learning. Existing approaches predominantly solve these challenges independently, while neglecting the fact that region detection and fine-grained feature learning are mutually correlated and thus can reinforce each other. In this paper, we propose a novel recurrent attention convolutional neural network (RA-CNN) which recursively learns discriminative region attention and region-based feature representation at multiple scales in a mutual reinforced way. The learning at each scale consists of a classification sub-network and an attention proposal sub-network (APN). The APN starts from full images, and iteratively generates region attention from coarse to fine by taking previous prediction as a reference, while the finer scale network takes as input an amplified attended region from previous scale in a recurrent way. The proposed RA-CNN is optimized by an intra-scale classification loss and an inter-scale ranking loss, to mutually learn accurate region attention and fine-grained representation. RA-CNN does not need bounding box/part annotations and can be trained end-to-end. We conduct comprehensive experiments and show that RA-CNN achieves the best performance in three fine-grained tasks, with relative accuracy gains of 3.3%, 3.7%, 3.8%, on CUB Birds, Stanford Dogs and Stanford Cars, respectively.

This paper focuses on indoor semantic segmentation using RGB-D data. Although the commonly used deconvolution networks (DeconvNet) have achieved impressive results on this task, we find there is still room for improvements in two aspects. One is about the boundary segmentation. DeconvNet aggregates large context to predict the label of each pixel, inherently limiting the segmentation precision of object boundaries. The other is about RGB-D fusion. Recent state-of-the-art methods generally fuse RGB and depth networks with equal-weight score fusion, regardless of the varying contributions of the two modalities on delineating different categories in different scenes. To address the two problems, we first propose a locality-sensitive DeconvNet (LS-DeconvNet) to refine the boundary segmentation over each modality. LS-DeconvNet incorporates locally visual and geometric cues from the raw RGB-D data into each DeconvNet, which is able to learn to upsample the coarse convolutional maps with large context whilst recovering sharp object boundaries. Towards RGB-D fusion, we introduce a gated fusion layer to effectively combine the two LS-DeconvNets. This layer can learn to adjust the contributions of RGB and depth over each pixel for high-performance object recognition. Experiments on the large-scale SUN RGB-D dataset and the popular NYU-Depth v2 dataset show that our approach achieves new state-of-the-art results for RGB-D indoor semantic segmentation.

We propose local binary convolution (LBC), an efficient alternative to convolutional layers in standard convolutional neural networks (CNN). The design principles of LBC are motivated by local binary patterns (LBP). The LBC layer comprises of a set of fixed sparse pre-defined binary convolutional filters that are not updated during the training process, a non-linear activation function and a set of learnable linear weights. The linear weights combine the activated filter responses to approximate the corresponding activated filter responses of a standard convolutional layer. The LBC layer affords significant parameter savings, 9x to 169x in the number of learnable parameters compared to a standard convolutional layer. Furthermore, the sparse and binary nature of the weights also results in up to 9x to 169x savings in model size compared to a standard convolutional layer. We demonstrate both theoretically and experimentally that our local binary convolution layer is a good approximation of a standard convolutional layer. Empirically, CNNs with LBC layers, called local binary convolutional neural networks (LBCNN), achieves performance parity with regular CNNs on a range of visual datasets (MNIST, SVHN, CIFAR-10, and ImageNet) while enjoying significant computational savings.

The goal of this work is to recognise phrases and sentences being spoken by a talking face, with or without the audio. Unlike previous works that have focussed on recognising a limited number of words or phrases, we tackle lip reading as an open-world problem - unconstrained natural language sentences, and in the wild videos. Our key contributions are: (1) a 'Watch, Listen, Attend and Spell' (WLAS) network that learns to transcribe videos of mouth motion to characters; (2) a curriculum learning strategy to accelerate training and to reduce overfitting; (3) a 'Lip Reading Sentences' (LRS) dataset for visual speech recognition, consisting of over 100,000 natural sentences from British television. The WLAS model trained on the LRS dataset surpasses the performance of all previous work on standard lip reading benchmark datasets, often by a significant margin. This lip reading performance beats a professional lip reader on videos from BBC television, and we also demonstrate that if audio is available, then visual information helps to improve speech recognition performance.

Linking two data sources is a basic building block in numerous computer vision problems. Canonical Correlation Analysis (CCA) achieves this by utilizing a linear optimizer in order to maximize the correlation between the two views. Recent work makes use of non-linear models, including deep learning techniques, that optimize the CCA loss in some feature space. In this paper, we introduce a novel, bi-directional neural network architecture for the task of matching vectors from two data sources. Our approach employs two tied neural network channels that project the two views into a common, maximally correlated space using the Euclidean loss. We show a direct link between the correlation-based loss and Euclidean loss, enabling the use of Euclidean loss for correlation maximization. To overcome common Euclidean regression optimization problems, we modify well-known techniques to our problem, including batch normalization and dropout. We show state of the art results on a number of computer vision matching tasks including MNIST image matching and sentence-image matching on the Flickr8k, Flickr30k and COCO datasets.

In this paper, we study learning visual classifiers from unstructured text description at part precision with no training images. We show that visual text terms can be encouraged to attend to its relevant parts, while image connections to non-visual text terms vanishes without any supervision. This learning process enables terms like "peak" to be linked to parts like only head for instance , while non-visual terms like "migrate" not to affect classifier prediction without part-text annotation. Images are encoded by a part-based CNN that detect bird parts and learn part-specific learning representation. Part-based visual classifiers are predicted from text descriptions of unseen visual classifiers to facilitate classification without training images (also known as zero-shot recognition ). We performed our experiments on CUB200 dataset and improves the zero-shot recognition results from 34.2% to 44.0%. We also created a large scale benchmark on 404 North American Bird Images with text descriptions, where we also showed that our method outperforming existing methods.

In this paper, we take advantage of the clear texture structure of the epipolar plane image (EPI) in the light field data and model the problem of light field reconstruction from a sparse set of views as a CNN-based angular detail restoration on EPI. We indicate that one of the main challenges in sparsely sampled light field reconstruction is the information asymmetry between the spatial and angular domain, where the detail portion in the angular domain is damaged by undersampling. To balance the spatial and angular information, the spatial high frequency components of an EPI is removed using EPI blur, before feeding to the network. Finally, a non-blind deblur operation is used to recover the spatial detail suppressed by the EPI blur. We evaluate our approach on several datasets including synthetic scenes, real-world scenes and challenging microscope light field data. We demonstrate the high performance and robustness of the proposed framework compared with the state-of-the-arts algorithms. We also show a further application for depth enhancement by using the reconstructed light field.